DeepSeek V4 dropped this morning — April 24 — and the coverage is already full of parameter counts and benchmark charts. That's fine. But the thing most teams are going to care about three months from now isn't the model size. It's a single number buried in the release notes: one million tokens of context.
That sounds abstract until you do the math. One million tokens is roughly 750,000 words. That's your entire internal wiki, your full grant history, your last two years of meeting notes, your complete project documentation for a mid-size initiative. All of it. In a single prompt.
And doing that now costs about fourteen cents.
What just happened
V4 ships in two tiers. V4-Flash ($0.14/M input, $0.28/M output) is the workhorse — fast, cheap, one million token window. V4-Pro ($1.74/M input) runs the full 1.6 trillion parameter model with 49 billion activated per token, same context window, for when you need serious reasoning or code generation at scale. Both are open source under permissive licensing, weights published on Hugging Face, live API at api.deepseek.com right now.
The API accepts both OpenAI and Anthropic-formatted requests, which means if you're already using a routing layer, this is a config change, not a refactor.
The tax you've been paying without realizing it
For the past two years, the standard advice for "talk to your documents with AI" has been: build a RAG pipeline.
RAG — Retrieval-Augmented Generation — means you chunk your documents into pieces, run them through an embeddings model, store the vectors in a database, then at query time retrieve the most relevant chunks and feed those to the LLM. The LLM only ever sees a slice of your knowledge base per question, not the whole thing.
This works. But it introduces a whole layer of moving parts: an embeddings service, a vector database (Pinecone, Weaviate, Qdrant, or something you're hosting yourself), chunking logic, retrieval tuning, a reranker if you want better results. Realistically, setting this up properly for a 10-person team takes several days of engineering time. Maintaining it takes ongoing attention. Doing it badly — and chunking strategies are genuinely tricky — means your AI answers questions based on fragments, misses context, and occasionally hallucinates connections between unrelated pieces.
The hidden assumption underneath all of this was: context windows are small and expensive, so you have to be selective about what the model sees.
That assumption just expired.
Three things you can actually do this week
1. Feed your entire document corpus into a conversation
If your organization's total written knowledge fits under about 1,500 pages — policies, procedures, past reports, meeting notes, program documentation — you can now load all of it into a single V4-Flash session and ask questions against the whole thing.
No chunking. No vector database. No retrieval misses. The model sees everything simultaneously and can reason across documents that a RAG pipeline would never retrieve together.
Cost to process the whole corpus each time: somewhere between $0.07 and $0.20, depending on how much text you have. At that price, you run it on every question without thinking about it.
For a 15-person NGO whose institutional knowledge lives in a shared drive, this is the difference between "we need a developer to build this" and "we set this up on a Wednesday afternoon."
2. Analyze an entire RFP or government tender in one pass
A typical government procurement document runs 80 to 200 pages. The standard approach is to either read it yourself (hours), pay someone to summarize it (expensive), or use an AI tool that chunks it and loses the cross-references between sections (frustrating).
With a 1M context window, you paste the whole thing in — or point an API call at the PDF — and ask for a structured breakdown: eligibility criteria, evaluation weighting, required certifications, red flags, how this maps to your org's existing capabilities. The model sees the full document and can catch when a clause in section 4 quietly contradicts an eligibility statement in section 12.
This is genuinely useful for any org that responds to tenders or applies for grants regularly. It doesn't replace a thorough review, but it collapses the first-pass analysis from three hours to fifteen minutes.
3. Cross-document reasoning you couldn't do before
Here's the thing about chunked retrieval that doesn't get talked about enough: it's bad at questions that require synthesizing across many documents simultaneously.
"What decisions have we made about housing policy across all our program reports from the last three years?" A RAG pipeline retrieves the five or six most semantically similar chunks and answers from those. It misses things. It can't count. It loses track of context between documents.
With full-context ingestion, you can load all three years of reports and get an answer that actually reflects the full picture. For organizations doing institutional research, policy analysis, or program evaluation, this is a qualitative shift in what's possible without specialist tooling.
The honest caveats
This approach breaks down once your document base grows beyond what fits comfortably in 1M tokens — roughly 1,500 to 2,000 pages depending on density. Larger organizations or orgs with extensive archives will still need RAG or a proper hybrid approach. The context window isn't infinite.
And the data sovereignty question doesn't go away. If you're feeding client files, health records, personal information, or anything regulated into this API, you're sending that data to DeepSeek's servers in China. For those workloads, you need either a self-hosted deployment (the weights are on Hugging Face, so it's technically possible, but you need serious hardware) or a provider with Canadian or EU data residency. Don't let the cheap API price make you skip that assessment.
For non-sensitive workloads — public documents, internal procedures, published research, grant-eligible program documentation — there's no real barrier to getting started today.
Why this matters beyond the model specs
The pattern here is worth noticing. Every few months, something that required a dedicated technical build — a vector database, an embedding pipeline, a custom retrieval layer — gets absorbed into the baseline capability of a cheap model. The tooling tax for building useful AI workflows keeps dropping.
For small orgs, this is asymmetrically good news. A 5-person team has the same access to a 1M-token context window as a 500-person company. The difference is who actually figures out how to use it before their sector does.
RAG isn't dead for complex use cases. But for the document-heavy workflows that most NGOs, public sector teams, and small businesses actually have? The "you need to build a retrieval system" default answer just got a lot more questionable.
The API key takes two minutes to set up. The use case you've been putting off because it felt like too much infrastructure might be simpler than you thought.
If you want to figure out which of your document workflows are a good fit for this approach — and which ones need more careful data handling — we can scope that quickly. Get in touch.